Mnemosyne OS · XPACEGEMS LLC
We build an AI assistant that remembers you, on machines you control.
Mnemosyne OS keeps your documents, notes and conversations as long-term memory on your own hardware, and every model you connect answers from it. It reads the folders you open to it and nothing else. Vaults can be encrypted with AES-256 at rest once you turn it on, under a key only you hold. The cloud is optional: your key, your Vertex project or your own server.
Local-first · Optional AES-256 at rest · 77.1% LongMemEval-M · The human governs
Windows, macOS and Linux.
What XPACEGEMS builds
A personal assistant, its memory, and a desk for your agents.
XPACEGEMS LLC is an independent AI lab in Miami. Its product, Mnemosyne OS, is a desktop application for Windows, macOS and Linux. It turns a person's files and conversations into a memory that any AI model can draw on.
The same application supervises the AI agents a person runs. Coding agents and the Hermes assistant report to it, and a command that could do damage waits for a human decision.
Sources and verification
Every answer shows where it comes from.
While the answer is written, the notes and documents it draws on appear beside it. Turn on verification and a second pass compares each sentence with those sources. The sentences no source supports are flagged.
Our research
The relationship layer, and the half we have not finished.
Memory
Vaults per domain, retrieval by meaning and by exact words, sources shown under every answer. This half has a public benchmark: 77.1% on unseen LongMemEval-M questions, with every verdict published.
How it behaves with a person
An assistant that knows someone still has to decide how to talk to them. Ours questions plans instead of approving them, which surprises people the first time. No benchmark captures this, so we study it from the conversations people decide to send us.
Did the way it answers surprise you? Write to the lab: dev@mnemosyne-os.io
In version 1.8.0
What it does today
- LaTeX
herdr- Headless MCP
- VS Code
- Cursor
- Claude Code
- Hermes
- Telegram
-
Agents on one desk
Each coding agent you run in herdr gets its own card, which tells you if it is working or waiting on you. The Hermes assistant is set up with a single click and replies on Telegram.
-
Gesture and voice control“JARVIS mode”
Speak to the orb and it carries out commands, such as opening your agenda. With a webcam you move windows by hand. Speech and hand tracking stay on the machine.
-
Memory cartridges
Public knowledge fetched from official sources into your memory: laws, medicine, science, world history, software vulnerabilities, trademark registers and languages. The chat quotes each text with its source and its date.
-
Answers you can check
A second pass compares each sentence of an answer with the sources it used, and flags the sentences no source supports.
-
Offline when it matters
With a local model you can unplug the internet cable. A lawyer or a journalist questions a sensitive file, and the file stays on the machine.
-
For developers
An MCP server connects Claude Code, Cursor and other agents to the same vaults. An SDK lets you build your own cartridges.
Where you come from
Enterprise
Years of documents and conversations, searchable on your own hardware. You decide what is kept.
Available today Read the page → 02Journalism
Your whole investigation file, searchable: PDFs, interviews, captures, notes. Your sources stay on your machine. It does not write your articles.
Available today Read the page → 03Research labs
77.1% on LongMemEval-M, every verdict published. And an offer: a memory engine for your discipline, built on your corpus.
Open collaboration Read the page → 04Public sector
Your archives stay in your jurisdiction, on your hardware. No foreign cloud is imposed.
Read the page → 05Banking & finance
All of a client’s context in one place, on the bank’s own servers.
Read the page → 06Robotics, the frontier
A robot must keep its memory on board. Latency breaks its reflexes, and a network drop must not make it forget. Memory on board, encrypted, offline.
Read the page →Pick your sector. Each page says what Mnemosyne OS does for you.
A score you can check.
Our memory engine scores 77.1% on LongMemEval-M, a public benchmark for long-term conversational memory, on 48 questions it had never seen, with the official LongMemEval judge. The protocol was published before the run, and every answer is published with its verdict. On the 48 questions used to tune the engine, the strict judge gives 85.4%. Hybrid mode: embeddings and search run on the machine, a cloud model (gemini-3.8-flash) writes the answers.
The grader, the verdicts and the method are published. You can redo the math yourself. The engine stays private. Our failures are documented next to our successes.
LongMemEval-M · 77.1% on unseen questions · official judge · protocol published before the run · misses documented
The benchmark itself is public and third-party: LongMemEval on GitHub →
The full dossier (numbers, protocol, receipts) lives on the product site: mnemosyne-os.io/benchmark →
The published figures
LongMemEval-M, full-haystack variant: about 480 distractor sessions per question. Two sets of 48 questions and two judges. They sit side by side, not as a progression.
Unseen questions
77.1%
37/48 on a pre-registered holdout, official LongMemEval judge
Tuning questions
85.4%
41/48 on the questions used to tune the engine, strict judge
Speed
Write and recall times
Measured on one machine, an Intel Core Ultra 7 155H, on the processor alone, no GPU, core-engine 1.2.0. We ran the measurement twice and show the slower run. These are our own measurements, not a graded benchmark.
Write
0.10ms
p50 into the vault. The embedding before it is not counted
Vector recall
20ms
p50, top 10, warm cache, lexical channel off, 10,200 memories
Hybrid recall
108ms
What the chat actually runs: BM25 fused into the ranking
The hybrid recall is the default path since August 17. It is the one behind these figures. It ranks 200 candidates and adds a BM25 pass, so it costs 5.3× a vector-only recall. Both answer from RAM. While the cache fills, a vector recall reads the whole index in 145 ms.
Mnemosyne Labs
Our first work is published: a technical whitepaper on the architecture and the benchmark audit kit, both with a DOI and attached to an ORCID. The lab notes say what we measured, and what turned out to be wrong.
Published Technical whitepaper, DOI 10.5281/zenodo.21728283Benchmark audit kit, DOI 10.5281/zenodo.21727140ORCID 0009-0009-1087-3917
References The research lineage we build on
Two pieces of public research are the basis of the architecture. We extend the first. We measure ourselves on the second.
-
The “LLM as an operating system” idea: a context managed the way an OS manages memory. Mnemosyne OS extends it with a service that runs continuously, isolates agents and keeps everything on disk, instead of a library loaded at each session.
arXiv:2310.08560 · MemGPT · Packer et al. · UC Berkeley · 2023
-
An independent academic benchmark for long-term memory, not a metric of our own. Our 77.1% is measured on it.
arXiv:2410.10813 · LongMemEval · Wu et al. · ICLR 2025
Contact
For anything that fits no section. A person reads every message.